QASSIT at SemEval-2016 Task 13: On the integration of Semantic Vectors in Pretopological Spaces for Lexical Taxonomy Acquisition

Proceedings of SemEval-2016, pages 1315–1319, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics QASSIT at SemEval-2016 Task 13: On the Integration of Semantic Vectors in Pretopological Spaces for Lexical Taxonomy Acquisition Guillaume Cleuziou Universit´e d’Orl´eans, INSA Centre Val de Loire, LIFO EA 4022, France cleuziou@univ-orleans.fr Jose G. Moreno LIMSI, CNRS, Universit´e Paris-Saclay, F-91405 Orsay moreno@limsi.fr Abstract This paper presents our participation to the Se- mEval “Task 13: Taxonomy Extraction Evalu- ation (TExEval-2)” (Bordea et al., 2016). This year, we propose the combination of recent se- mantic vectors representation into a method- ology for semisupervised and auto-supervised acquisition of lexical taxonomies from raw texts. In our proposal, first similarities be- tween concepts are calculated using semantic vectors, then a pretopological space is defined from which a preliminary structure is con- structed. Finally, a genetic algorithm is used to optimize two different functions, the qual- ity of the added relationships in the taxonomy and the quality of the structure. Experiments show that our proposal has a competitive per- formance when compared with the other par- ticipants achieving the second position in the general rank. 1 Introduction The task of automatic taxonomy extraction consists in the generation of hierarchical relations between pairs of terms from a given initial set of terms. In this paper, we describe the second participation of QASSIT team to the “Taxonomy Extraction Eval- uation (TExEval)” task. In this opportunity, we were interested in re-examinating our proposed al- gorithm (Cleuziou and Dias, 2015)(Cleuziou et al., 2015), but with an improved range of information as input. Our algorithm is based on Pretopologi- cal Spaces combined with patterns and collocations counts. Patterns are a useful way to capture knowl- edge form huge corpus (Hearst, 1992), but its per- formance could be influence by the size of the cor- pus or the use of a fixed vocabulary. This is a main drawback of the previous version of the al- gorithm (Cleuziou et al., 2015), the requirement of non-zero mentions in a corpus of the input terms in the specific chosen patterns. For example, to an optimal performance our algorithm requires that the term “computational biology” appears at least once in one of the used patters. However, when search- ing in wikipedia with the query “computational bi- ology is a ...” then we obtain results such as “PLOS Computational Biology is a peer-reviewed compu- tational biology journal”, “The Journal of Compu- tational Biology is a monthly peer-reviewed scien- tific journal”, “Cancer computational biology is a field that aims to determine the future mutations in” and “Computational biology is a specialized do- main that often requires knowledge of computer pro- gramming”1. The first three results include the de- sired term as a part of another more specific term (in this case journal titles) and only the last result could be considered as relevant information. More- over, neither “domain” or “computer programming” are part of the input terms for the taxonomy. This sit- uation makes difficult the direct use of patterns with our …

Content truncated to 500 words